RES-71: Scientific Engine Qualification Gate

docs/qualification/RES71-COVERAGE-MATRIX.md

@@ -0,0 +1,40 @@
+# RES-71 V2 coverage matrix
+
+`build_coverage_matrix()` returns the machine-readable version. Every row has
+authoritative surfaces, registered operations (where applicable), unresolved
+capabilities, provenance boundary, comparability boundary, claim boundary,
+authority references and test paths. Validation keys supplied and canonical
+rows by `domain` and requires exact frozen-dataclass equality before preserving
+the existing domain and test-path checks.
+
+| V2 domain | Status | Explicit boundary / unresolved surface |
+|---|---|---|
+| Population/source authority | QUALIFIED | Ambiguous/noncanonical populations quarantine; indirect evidence cannot mint target priors. |
+| Football world/context | QUALIFIED | Match-day relation requires an identified target match; no nearest-match inference. |
+| Ingestion/qualification | QUALIFIED | Raw bytes, license, schema, mapping, canonical artifact and replay lineage remain bound. |
+| External load / GNSS / optical | QUALIFIED_WITH_EXPLICIT_DEFERRED | Hidden vendor algorithms and unresolved threshold/sampling semantics are observations/refusals, not reconstructed equations. |
+| CMJ | QUALIFIED_WITH_EXPLICIT_DEFERRED | CMJ RFD and COM-displacement jump height remain deferred; estimator/event/phase identities remain distinct. |
+| Strength / IMTP / VBT | QUALIFIED_WITH_EXPLICIT_DEFERRED | VBT MPV and terminal-velocity estimated 1RM are represented/refused; velocity loss is not fatigue. |
+| Sprint / max velocity / COD / RSA / 30–15 IFT | QUALIFIED_WITH_EXPLICIT_DEFERRED | Acceleration, sustained maximum, 505 asymmetry and VIFT relabelling remain refused/deferred. |
+| DJ / bench throw / medicine-ball throw | QUALIFIED_WITH_EXPLICIT_DEFERRED | DJ JH/CT differs from FT/CT; generic BPT/MBT power and norms are refused. |
+| Longitudinal statistics / reliability / error | QUALIFIED_WITH_EXPLICIT_DEFERRED | Deferred ICC/SEM/MDC/CI/BA/repeated-measures methods are explicit; no generic meaningful-change authority. |
+| Cross-source comparability | QUALIFIED_WITH_EXPLICIT_DEFERRED | Exact identities, context and registry hashes are required; no transitive or caller-minted bridge. |
+| Analysis capability | QUALIFIED_WITH_EXPLICIT_DEFERRED | Missing prerequisites, wrong level, pseudoreplication and deferred between/cross-test models refuse. |
+| Claim authority | QUALIFIED_WITH_EXPLICIT_DEFERRED | Ladder/causal level is enforced; practical, readiness, fatigue, injury and causal escalation refuse. |
+
+“Qualified with explicit deferred” means the domain is covered by an
+authoritative implementation/refusal contract; it does not claim that every
+published method is implemented.
+
+## Required upstream relationships
+
+```text
+population/source → football context → canonical record/lineage
+       ↓                    ↓                    ↓
+measurement identity → deterministic operation → statistical analysis
+       ↓                    ↓                    ↓
+comparability/evidence → claim authority → bounded result or refusal
+```
+
+No arrow permits a label, unit, correlation, provider output or LM proposal to
+skip identity, provenance, comparability, analysis or claim prerequisites.

docs/qualification/RES71-GATE-RECEIPT.json

@@ -0,0 +1,26 @@
+{
+  "MISSION": "RES-71-SCIENTIFIC-ENGINE-GATE-001",
+  "BASE_MAIN": "7508a9025759c2863d163e09b22f325494828602",
+  "QUALIFIED_CONTENT_HEAD": "0a51127628f1bfc1f0b89064bf92d7fc2703ff39",
+  "STATUS": "PASS",
+  "REGISTERED_OPERATION_INVENTORY": "PASS",
+  "COVERAGE_MATRIX": "PASS",
+  "UNRESOLVED_COMPUTATION_INVENTORY": "PASS",
+  "REFERENCE_CASE_VALIDATION": "PASS",
+  "RUNTIME_COUNTS": {
+    "registered_operations": 100,
+    "implemented": 81,
+    "historical_replay_only": 1,
+    "represented_but_do_not_compute": 7,
+    "deferred": 8,
+    "rejected": 3,
+    "unresolved_capabilities": 23,
+    "coverage_domains": 12,
+    "verifier_reference_cases": 12,
+    "verifier_reference_digest": "sha256:d29d84699b7cf70c2d409d370c5ffd6c7ad7cd704375b14b541527a95fa385e5"
+  },
+  "SERIALIZATION_VERSION": 3,
+  "MODEL_INFERENCE": "NOT_IMPLEMENTED",
+  "GPU_WORK": "NOT_IMPLEMENTED",
+  "SCIENTIFIC_ENGINE_GATE": "PASS"
+}

docs/qualification/RES71-OPERATION-INVENTORY.md

@@ -0,0 +1,72 @@
+# RES-71 registered-operation inventory
+
+The authoritative inventory is the immutable tuple returned by
+`dynamislm.qualification.build_registered_operation_inventory()`.
+`validate_registered_operation_inventory()` discovers every
+`RegistryReference` whose object type is `registered-operation`, keys supplied
+and canonical rows by `operation_id`, and requires exact frozen-dataclass
+equality before route/path/runtime checks. A newly exposed operation or a
+metadata substitution therefore fails qualification until its contract is
+reviewed.
+
+Each row contains:
+
+```text
+operation_id
+label
+method_version
+scientific_family
+disposition
+implementation
+input_contract
+output_contract
+provenance_contract
+refusal_path
+test_coverage
+authority_references
+tolerance_contract
+```
+
+## Inventory totals
+
+| Family | Registered identities | Implemented | Historical replay | Represented only | Deferred | Rejected |
+|---|---:|---:|---:|---:|---:|---:|
+| Population/source | 2 | 2 | 0 | 0 | 0 | 0 |
+| Football context | 1 | 1 | 0 | 0 | 0 | 0 |
+| Ingestion/record | 3 | 3 | 0 | 0 | 0 | 0 |
+| External load | 3 | 3 | 0 | 0 | 0 | 0 |
+| CMJ | 25 | 24 | 1 | 0 | 0 | 0 |
+| Strength/IMTP/VBT | 14 | 12 | 0 | 1 | 1 | 0 |
+| Field testing | 14 | 12 | 0 | 1 | 1 | 0 |
+| DJ/BPT/MBT | 10 | 9 | 0 | 1 | 0 | 0 |
+| Longitudinal statistics | 27 | 14 | 0 | 4 | 6 | 3 |
+| Cross-source bridge | 1 | 1 | 0 | 0 | 0 | 0 |
+| **Total** | **100** | **81** | **1** | **7** | **8** | **3** |
+
+The count is a qualification assertion, not a replacement for the exact set
+comparison. The machine-readable entries carry the implementation symbol,
+input/output/provenance contract, refusal route, authority documents and test
+paths for each identity.
+
+## Authority mapping by family
+
+| Family | Primary implementation surface | Authority | Test evidence |
+|---|---|---|---|
+| Population/source | `dynamislm.population`, `dynamislm.ingestion` | RES-59/60/63 | `test_population.py`, `test_ingestion.py`, `test_res63_operational_authority.py` |
+| Football context | `dynamislm.football`, `dynamislm.longitudinal` | RES-61/62 | `test_football.py`, `test_longitudinal.py` |
+| External load | `dynamislm.external_load` | RES-64 | `test_external_load.py` |
+| CMJ | `dynamislm.measurement.cmj` | RES-34–50/65 | CMJ test family and RES-71 gold/refusal cases |
+| Strength | `dynamislm.measurement.strength` | RES-66 | `test_strength.py` |
+| Field testing | `dynamislm.measurement.field_testing` | RES-67 | `test_field_testing.py` |
+| DJ/BPT/MBT | family-specific RES-68 packages | RES-68 | `test_explosive_test_families.py` |
+| Longitudinal statistics | `dynamislm.longitudinal.statistics` | RES-69 | `test_longitudinal_statistics.py` |
+| Comparability | `dynamislm.comparability` | RES-70 | RES-70 comparability/bridge/adversarial tests |
+
+## Completeness rule
+
+Implemented entries must resolve to an importable implementation symbol and
+have test paths in the repository. Historical replay entries resolve to the
+historical method identity and cannot mint the current result. Every other
+disposition has an explicit refusal/representation path. Caller-supplied
+formulas, thresholds, operation identities and comparability overrides are not
+accepted by this inventory.

docs/qualification/RES71-PROVENANCE-SERIALIZATION.md

@@ -0,0 +1,49 @@
+# RES-71 provenance and serialization qualification
+
+## Identity and lineage invariants
+
+The qualification accepts an output only when the owning family contract keeps
+the following separable:
+
+```text
+ObservationContext + MeasurementIdentity + MeasurementResult + Provenance
+```
+
+Derived results must preserve source observation/artifact IDs, acquisition and
+processing lineage, method/registry/software version, calculation-changing
+parameters, units, value origin, quality/uncertainty status and evidence or
+decision references. A method identity is not a method instance, and a provider
+output is not silently converted into a DynamisLM direct measurement.
+
+Reprocessing follows the sealed invariant:
+
+```text
+raw R + method v1 → D1
+raw R + method v2 → D2
+D2 does not overwrite D1
+```
+
+The RES-71 inventory requires provenance contracts for all 100 operations. The
+RES-70 authority revalidates exact observation hashes and registry hashes at
+comparability, analysis and claim intake. RES-63 remains the authority for raw
+artifact, canonical-artifact and qualification lineage.
+
+## Serialization
+
+- Serialization V3 remains unchanged.
+- Historical CMJ V1 flight-time results are replay-only and remain distinct
+  from the current V2 method/version.
+- Qualification contracts are registered with the existing canonical serializer
+  and round-trip through `canonical_json` / `from_canonical_json`.
+- The verifier manifest has a deterministic digest:
+  `sha256:9807b6e0be63abd44135bc855a97d37775ef09763d25c0f7025f4395ab673af6`.
+- Floating reference values carry explicit absolute/relative tolerance fields;
+  nonfinite values are rejected by the typed reference contract and owning
+  numerical operations.
+
+## Evidence
+
+The RES-71 tests cover serialization round-trip, canonical digest stability,
+independent numerical gold values, finite/domain refusal, source/method
+identity collision, historical/current method separation and safe refusal
+states. Full repository QA is required again at the exact final head.

docs/qualification/RES71-SCIENTIFIC-ENGINE-QUALIFICATION.md

@@ -0,0 +1,114 @@
+# RES-71 scientific-engine qualification report
+
+Mission: `RES-71-SCIENTIFIC-ENGINE-GATE-001`
+
+Qualified base: `7508a9025759c2863d163e09b22f325494828602`
+
+Scope: deterministic V2 scientific truth layer only.
+
+## Executive result
+
+The qualification harness independently audits the merged RES-59 and
+RES-63–70 implementation surface. It discovers 100 unique
+`registered-operation` identities at runtime and requires a reviewed inventory
+entry for every one. The current contract contains 81 implemented operations,
+one historical replay-only identity, seven represented-but-not-computed
+identities, eight deferred identities and three rejected identities.
+
+The deferred and rejected surfaces are explicit refusal authority. They do not
+count as missing gate coverage: unsupported methods are represented with a
+reason, refusal path, safe description and tests.
+
+The formal gate decision is sealed in `RES71-GATE-RECEIPT.json` after final QA.
+
+## RES-71 review-fix-002 qualification-integrity hardening
+
+Review-fix entry head: `dc29909a13fbb27d463632f2cd5c4f239c4cd07f`
+
+Qualified content head: `0a51127628f1bfc1f0b89064bf92d7fc2703ff39`
+
+The review fix canonically binds registered-operation rows by `operation_id`,
+unresolved-computation rows by `capability`, and coverage rows by `domain`.
+Supplied rows must equal freshly built canonical frozen records before route,
+path or refusal-runtime checks. The receipt comparison is recursively
+type-strict, including nested `RUNTIME_COUNTS` values.
+
+`RES71-GATE-RECEIPT.json` is deliberately narrow and runtime-authoritative.
+It contains only identity fields and evidence derived by
+`validate_gate_receipt()`. The former checklist assertions
+(`IDENTITY_PROVENANCE`, `NUMERICAL_QUALIFICATION`, `SCIENTIFIC_BOUNDARIES`,
+`DATASET_COMPATIBILITY`, `VERIFIER_REFERENCE_INTERFACE`,
+`IMPLICIT_LM_ARITHMETIC`, `UNREGISTERED_ACCEPTED_OPERATION`,
+`PROVENANCE_GAPS`, `CLAIM_AUTHORITY_BYPASS`, and `SCIENTIFIC_BLOCKERS`) are
+narrative qualification documentation, not unchecked fields in the sealed
+runtime receipt. CI results and test counts are handoff evidence outside that
+receipt.
+
+## Entry and upstream authority
+
+The entry checks were performed before mutation:
+
+| Check | Result |
+|---|---|
+| One worktree | PASS |
+| Branch | `work/res-71-scientific-engine-qualification-gate` |
+| `HEAD` | `54c2c21171e951081bda4bb23c88fa06986a2d79` |
+| `origin/main` | `7508a9025759c2863d163e09b22f325494828602` |
+| Worktree | clean at entry |
+| Linear RES-71 | In Progress; no LM/GPU work authorized |
+
+The qualification treats the sealed V2 constitution, measurement/provenance
+architecture, RES-60–70 decision records/receipts, and merged repository state
+as upstream authority. It does not alter their scientific formulas, thresholds,
+identities or expected fixtures.
+
+## Qualification components
+
+| Component | Evidence | Result |
+|---|---|---|
+| Registered-operation inventory | `build_registered_operation_inventory()` equals live runtime registry | PASS |
+| V2 coverage matrix | 12 required scientific domains | PASS |
+| Unresolved computation inventory | 23 explicit refusal/representation entries | PASS |
+| Identity/provenance | typed records, source/method/version/lineage checks and adversarial cases | PASS |
+| Numerical integrity | independent gold values, finite/domain refusals, serialization/hash checks | PASS |
+| Scientific boundaries | same-label collision, causal/estimand/refusal cases and upstream adversarial suites | PASS |
+| Dataset compatibility | RES-63 canonical/quarantine authority and lineage coverage | PASS |
+| Verifier interface | 12 immutable reference cases plus canonical manifest/digest adapter | PASS |
+
+## Scientific boundary attacks
+
+The qualification suite explicitly covers:
+
+- same display label with a different measurand or method;
+- method identity versus method instance and historical V1 versus current V2;
+- within-athlete versus between-athlete estimands and pseudoreplication;
+- correlation/agreement and bridge requirements;
+- direct, derived, provider-derived and model-estimated value origin;
+- nonfinite values, invalid domains and missing method prerequisites;
+- numerical change versus measurement error and practical meaning;
+- measurement change versus readiness, fatigue or injury interpretation;
+- association versus causal evidence;
+- canonical target-population evidence versus indirect method evidence;
+- cross-source bridge and threshold abuse;
+- claim-authority escalation and safe observation descriptions after refusal.
+
+The existing RES-70 adversarial tests and new RES-71 cases are both run. A
+refusal blocks the unsupported claim while preserving the strongest safe
+description of the underlying observation.
+
+## Scope prohibition
+
+This qualification adds no model runtime, inference path, benchmark harness,
+GPU work, CPT/DAPT, SFT/PEFT or RLVR/GRPO implementation. The reference
+adapter emits deterministic case contracts only; it does not execute an LM.
+
+## Reproduction
+
+```text
+./scripts/ci.sh
+python scripts/res71_reference_cases.py --digest
+python scripts/res71_reference_cases.py --manifest
+```
+
+The exact final commit and component statuses are recorded in the formal gate
+receipt.

docs/qualification/RES71-UNRESOLVED-COMPUTATIONS.md

@@ -0,0 +1,45 @@
+# RES-71 unresolved-computation inventory
+
+Unsupported methods remain explicit and non-authoritative. The complete
+machine-readable set is returned by
+`build_unresolved_computation_inventory()`. Validation keys supplied and
+canonical rows by `capability` and requires exact frozen-dataclass equality
+before runtime refusal-route checks.
+
+## Registered but not computed
+
+| Operation / capability | Disposition | Refusal boundary |
+|---|---|---|
+| 505 asymmetry | REPRESENT_BUT_DO_NOT_COMPUTE | No frozen denominator, direction or sign convention. |
+| BPT mean propulsive velocity | REPRESENT_BUT_DO_NOT_COMPUTE | No generic acceleration/gravity/propulsive-boundary authority. |
+| Classical BA limits | REPRESENT_BUT_DO_NOT_COMPUTE | V1 exposes only the registered narrow method-comparison summary. |
+| ICC variants, SEM from ICC, MDC/SDC | REPRESENT_BUT_DO_NOT_COMPUTE | No registered reliability estimand/design authority. |
+| VBT mean propulsive velocity | REPRESENT_BUT_DO_NOT_COMPUTE | Concentric velocity cannot be relabelled as MPV. |
+| Estimated 1RM from load–velocity model | DEFERRED | Terminal-velocity applicability is not authorized across devices/designs. |
+| Sustained maximum sprint velocity | DEFERRED | No registered dwell/sustain estimator. |
+| Log/repeated-measures BA, confidence intervals, covariance propagation | DEFERRED | Estimator, design and uncertainty contract not sealed. |
+| Repeated-measures correlation and mixed effects | DEFERRED | Later analysis mission required; no numeric output is emitted. |
+| Generic meaningful change | REJECTED | No universal decision criterion is authorized. |
+| Generic SEM | REJECTED | No registered estimand or design. |
+| Readiness/fatigue/injury interpretation | REJECTED | Outside the scientific-engine authority boundary. |
+
+## No registered numeric operation
+
+| Capability | Deterministic refusal |
+|---|---|
+| CMJ RFD | `refuse_unregistered_cmj_rfd` |
+| Generic BPT load × velocity power | `calculate_bpt_load_times_velocity_power` |
+| MBT distance-as-power / protocol-independent norm | `calculate_mbt_distance_as_power`, `calculate_mbt_protocol_independent_normative_score` |
+| Sprint acceleration from split averages | `refuse_sprint_acceleration` |
+| VIFT relabelled as VO2max, MAS or maximum sprint speed | `refuse_vift_as_vo2max`, `refuse_vift_as_mas`, `refuse_vift_as_mss` |
+
+The VIFT relabelling routes are identity refusals, not missing-computation
+refusals: their expected class is `IDENTITY_UNRESOLVED` with
+`MEASURAND_MISMATCH` and `METRIC_DEFINITION_MISMATCH` reason codes. CMJ RFD
+expects the producer reason code `NO_REGISTERED_OPERATION`.
+
+Every refusal preserves the blocked claim, reason code(s), missing information
+and safe description where the owning family exposes them. No placeholder
+number is returned. This is the intended behavior for later LM verifiers:
+`COMPUTATION_NOT_REGISTERED` is a scored scientific outcome, not an error to
+be bypassed.